During the past decade, the robotics community has been focused on developing online autonomous driving algorithms that can establish safe robot motion planning. For instance, in scenarios where robots need to be deployed in unstructured environments, it is not feasible to specify all the rules, leading to the adoption of online autonomous driving algorithms. However, designing models for unstructured environments presents challenges in terms of accuracy and overcoming safety hazards posed by real-time dynamic obstacles, which are essential conditions for the commercialization of robots. To address this, the Robotics Society has been utilizing model-free deep reinforcement learning controllers to build robot platforms.
However, model-free deep reinforcement learning algorithms currently lack guarantees for safety during the exploration process, and the evidence for trusting the neural network models is limited. Therefore, this paper aims to propose two types of safety filters to be applied to the deep reinforcement learning controllers used in End-to-End autonomous driving algorithms. First, a safety controller that combines control barrier functions with deep reinforcement learning is validated in the scenario of Adaptive Cruise Control, a state-of-the-art driver assistance system. Second, in multi-objective autonomous driving scenarios, safety is verified using a model predictive filter. A multi-objective deep reinforcement learning controller is developed for collision avoidance and target reaching, and model predictive filter are combined in serial and parallel to handle different forms of multi-objective scenarios. Ultimately, after verifying the performance of the algorithm through a simulation based on a custom-made digital twin, there are plans to conduct a proof-of-concept experiment by equipping autonomous driving hardware with the astonishing Model Predictive Filter algorithm that is being proposed.
The research results demonstrate that incorporating safety filters into autonomous driving algorithms can enhance the safety of deep reinforcement learning by designing safety constraints based on vehicle dynamics, thus opening up new horizons in robotics research.